Performance of heterogeneous computing with graphics processing unit and many integrated core for hartree potential...

Area of Science:

• Computational Chemistry
• Materials Science
• High-Performance Computing

Background:

• Electronic structure calculations are crucial for understanding materials at the atomic level.
• Large-scale calculations demand significant computational resources, driving the need for efficient computing architectures.
• Heterogeneous computing, combining different processor types, offers a potential solution for accelerating these demanding tasks.

Purpose of the Study:

• To evaluate the performance of heterogeneous computing architectures, specifically combining CPUs, GPUs, and Many Integrated Core (MIC) processors.
• To assess the efficiency of a work-stealing scheduler for dynamic workload distribution in heterogeneous environments.
• To demonstrate the application of these methods in large-scale electronic structure calculations, using Hartree potentials of silver nanoparticles as a case study.

Main Methods:

• Direct integral method utilizing a sinc basis set for calculating Hartree potentials.
• Implementation of a work-stealing scheduler for balanced workload distribution across heterogeneous processors (CPUs, GPUs, MIC).
• Performance benchmarking of various configurations: 20×CPU, 1GPU, 20×CPU + 1GPU, and 20×CPU + 2GPU.

Main Results:

• The 20×CPU + 1GPU configuration was up to ~1.5 times faster than 1GPU and ~3.1 times faster than 20×CPU.
• The 20×CPU + 2GPU configuration demonstrated a ~4.3 times speedup compared to 20×CPU.
• CPU + MIC configurations showed lower-than-expected performance gains due to significant MIC initialization overhead.

Conclusions:

• Heterogeneous computing with GPUs offers substantial performance improvements for large-scale electronic structure calculations.
• The work-stealing scheduler effectively manages workload distribution in heterogeneous environments.
• While MIC processors have theoretical potential, their practical application in this context was hindered by initialization overhead, suggesting GPUs are currently more advantageous for this specific application.